Strategies for building robust prediction models using data unavailable at prediction time.

Area of Science:

• Medical Informatics
• Clinical Epidemiology
• Machine Learning in Healthcare

Background:

• Hospital-acquired infections (HAIs) lead to increased patient morbidity, mortality, and prolonged hospital stays.
• Current risk prediction models for HAIs often use pre- and intraoperative data, but their performance is surpassed by models using postoperative data.
• Postoperative data, though more predictive, is unavailable at the time of surgical risk assessment.

Purpose of the Study:

• To investigate if temporarily unavailable postoperative data (TUP) can be leveraged to enhance the performance of HAI risk prediction models.
• To determine the impact of incorporating TUP data on the accuracy of HAI risk assessment at the end of surgery.

Main Methods:

• Developed and evaluated 12 distinct methods, including logistic/linear regression and deep learning techniques.
• Explored various intermediate data representations to effectively incorporate TUP data into prediction models.
• Compared single-task and multi-task learning frameworks to account for the hierarchical nature of HAI outcomes.

Main Results:

• The integration of TUP data consistently improved predictive performance across all tested models.
• Baseline models not utilizing TUP data failed to achieve top performance levels.
• Model performance varied across different HAI outcomes, with intermediate representation complexity and label incorporation being key factors.

Conclusions:

• The utilization of TUP data provides a significant advantage in improving predictive performance for HAIs.
• The enhancement in predictive accuracy was observed irrespective of the complexity of the machine learning models employed.